def __init__(self, hparams):
super().__init__()
self.hidden_size = hparams.role_size * hparams.filler_size
self.temperature = hparams.temperature
self.zeroth_filler_cell_state = Parameter(
empty(1, self.hidden_size).zero_())
self.filler_cell = LSTMCell(hparams.embedding_dim, self.hidden_size)
self.filler_hidden_to_number = Linear(self.hidden_size,
hparams.filler_number,
bias=False)
self.filler_dictionary = Linear(hparams.filler_number,
hparams.filler_size,
bias=False)
self.zeroth_role_cell_state = Parameter(
empty(1, self.hidden_size).zero_())
self.role_cell = LSTMCell(hparams.embedding_dim, self.hidden_size)
self.role_hidden_to_number = Linear(self.hidden_size,
hparams.role_number,
bias=False)
self.role_dictionary = Linear(hparams.role_number,
hparams.role_size,
bias=False)
self.zeroth_hidden_state = Parameter(empty(self.hidden_size).zero_())
def get_pytorch_lstm(input_size, used_lstm):
"""Load in a PyTorch LSTM that is a copy of the currently used LSTM."""
lstm = PytorchLSTM(input_size, 1)
lstm.bias_hh[:] = tensor(zeros((4, )), dtype=float64)[:]
lstm.bias_ih[:] = tensor(used_lstm.bias, dtype=float64)[:]
lstm.weight_hh[:] = tensor(used_lstm.weight_hh, dtype=float64)[:]
lstm.weight_ih[:] = tensor(used_lstm.weight_xh, dtype=float64)[:]
return lstm
def __init__(self,
state_size,
num_actions,
act_lim=1,
batch_size=1,
hidden_size=128,
num_layers=2,
dropout=0.85):
""" Construct a multilayer LSTM that computes the action given the state
The agent will first decide which dimension to act on and then decide the numerical value of the aciton on that dimension
- shape of input state is given by state_size
- dimensions of the orthogonal action space is given by num_actions, whereas act_lim gives the numerical bound for action values
Note: the last action dimension is assumed to be discrete, meaning the agent "does nothing".
- hidden_size should match that of the encoding network (i.e. the size of the encoding layer)
"""
super(PolicyNet, self).__init__()
self.state_size = state_size
self.num_actions = num_actions
self.act_lim = act_lim
self.batch_size = batch_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.dropout = dropout
# Create multilayer LSTM cells
self.cell_list = nn.ModuleList()
self.cell_list.append(
LSTMCell(input_size=state_size, hidden_size=hidden_size))
for i in range(1, num_layers):
self.cell_list.append(
LSTMCell(input_size=hidden_size, hidden_size=hidden_size))
# Linear layer that decides the dimension the agent wants to act on.
# Return the logits to be used to construct a Categorical distribution
self.FC_decision = Linear(hidden_size, num_actions)
# Linear layer that computes the mean value of the agent's action on each dimension
self.FC_values_mean = Linear(hidden_size, num_actions)
# Linear layer that computes the log standard deviation of the agent's action on each dimension
self.FC_values_logstd = Linear(hidden_size, num_actions)
# Variables to store lists of hidden states and cell states at the end of each time step so as to be used as
# the input values to the next time step
# Reset to None at the start of each episode
self.h_list = None
self.c_list = None
def __init__(self, channels, h_g, h_l, std, hidden_size, num_classes,
learned_start):
"""
Initialize the recurrent attention model and its
different components.
Args
----
- g: size of the square patches in the glimpses extracted
by the retina.
- k: number of patches to extract per glimpse.
- s: scaling factor that controls the size of successive patches.
- c: number of num_channels in each image.
- h_g: hidden layer size of the fc layer for `phi`.
- h_l: hidden layer size of the fc layer for `l`.
- std: standard deviation of the Gaussian policy.
- hidden_size: hidden size of the rnn.
- num_classes: number of num_classes in the dataset.
- num_glimpses: number of glimpses to take per image,
i.e. number of BPTT steps.
"""
super(RecurrentAttention, self).__init__()
self.std = std
self.sensor = glimpse_network(h_g, h_l, learned_start, channels)
self.rnn = LSTMCell(256, hidden_size)
self.decision = decision_network(hidden_size, 2)
self.illuminator = illumination_network(hidden_size, channels, std)
self.classifier = action_network(hidden_size, num_classes)
def __init__(
self,
decoding_dim: int,
target_embedding_dim: int,
attention: Optional[Attention] = None,
bidirectional_input: bool = False,
) -> None:
super().__init__(
decoding_dim=decoding_dim,
target_embedding_dim=target_embedding_dim,
decodes_parallel=False,
)
# In this particular type of decoder output of previous step passes directly to the input of current step
# We also assume that decoder output dimensionality is equal to the encoder output dimensionality
decoder_input_dim = self.target_embedding_dim
# Attention mechanism applied to the encoder output for each step.
self._attention = attention
if self._attention:
# If using attention, a weighted average over encoder outputs will be concatenated
# to the previous target embedding to form the input to the decoder at each
# time step. encoder output dim will be same as decoding_dim
decoder_input_dim += decoding_dim
# We'll use an LSTM cell as the recurrent cell that produces a hidden state
# for the decoder at each time step.
self._decoder_cell = LSTMCell(decoder_input_dim, self.decoding_dim)
self._bidirectional_input = bidirectional_input
def __init__(self,
embed: nn.Embedding = None,
hidden_size: int = 200,
dropout: float = 0.1,
layer: str = "rcnn",
z_rnn_size: int = 30,
):
super(DependentLatentModel, self).__init__()
self.layer = layer
emb_size = embed.weight.shape[1]
enc_size = hidden_size * 2
self.embed_layer = nn.Sequential(embed, nn.Dropout(p=dropout))
self.enc_layer = get_encoder(layer, emb_size, hidden_size)
if layer == "rcnn":
self.z_cell = RCNNCell(enc_size + 1, z_rnn_size)
else:
self.z_cell = LSTMCell(enc_size + 1, z_rnn_size)
self.z_layer = KumaGate(enc_size + z_rnn_size)
self.z = None # z samples
self.z_dists = [] # z distribution(s)
self.report_params()
def __init__(self, arch: dict):
super().__init__()
self.ldim = arch['latent_dim']
self.defaultSteps = arch['std_T']
self.Cinit = torch.nn.Parameter(torch.FloatTensor(self.ldim))
torch.nn.init.normal_(self.Cinit)
self.Linit = torch.nn.Parameter(torch.FloatTensor(self.ldim))
torch.nn.init.normal_(self.Linit)
self.rec_block = arch['recurrent_block']
if self.rec_block == 'test':
# self.block = BiPartialTestBlock(self.ldim, self.ldim, 2 * self.ldim, 2)
Ldim, Cdim, Hdim, Dpth = self.ldim, self.ldim, 2 * self.ldim, 2
self.Cmsg = batchMLP(Cdim, Hdim, Cdim, Dpth, False)
self.Lmsg = batchMLP(Ldim, Hdim, Ldim, Dpth, False)
self.Cu = batchMLP(Cdim * 2, Hdim, Cdim, Dpth, False)
self.Lu = batchMLP(Ldim * 3, Hdim, Ldim, Dpth, False)
elif self.rec_block in ['std_lstm', 'ln_lstm', 'gru']:
Ldim, Cdim, Hdim, Dpth = self.ldim, self.ldim, self.ldim, 4
self.Cmsg = batchMLP(Cdim, Hdim, Cdim, Dpth, False)
self.Lmsg = batchMLP(Ldim, Hdim, Ldim, Dpth, False)
if self.rec_block == 'std_lstm':
self.Cu = LSTMCell(self.ldim, self.ldim, True)
self.Lu = LSTMCell(self.ldim * 2, self.ldim, True)
elif self.rec_block == 'ln_lstm':
self.Cu = ln_LSTMCell(self.ldim, self.ldim, True)
self.Lu = ln_LSTMCell(self.ldim * 2, self.ldim, True)
elif self.rec_block == 'gru':
self.Cu = GRUCell(self.ldim, self.ldim, True)
self.Lu = GRUCell(self.ldim * 2, self.ldim, True)
self.cl = arch['classifier']
if self.cl == 'NeuroSAT':
self.Lvote = batchMLP(self.ldim, 2 * self.ldim, 1, 2, False)
elif self.cl == 'CircuitSAT-like':
self.tnormf = arch['tnorm']
if 'tnorm_train' in arch:
self.train_tnorm = arch['tnorm_train']
else:
self.train_tnorm = self.tnormf
self.tnorm_tmp = arch['tnorm_temperature']
self.train_temp = arch['temp_train']
self.test_temp = arch['temp_test']
self.Lvote = batchMLP(self.ldim, 2 * self.ldim, 1, 2, False)
def __init__(self, arch=None):
"""
:param arch: dictionary, for overriding default architecture
"""
nn.Module.__init__(self)
self.arch = deepcopy(default_arch)
if arch is not None:
self.arch.update(arch)
self.T = self.arch.max_steps
self.reinforce_weight = 0.0
# 4: where + pres
lstm_input_size = self.arch.input_size + self.arch.z_what_size + 4
self.lstm_cell = LSTMCell(lstm_input_size, self.arch.lstm_hidden_size)
# predict z_where, z_pres from h
self.predict = Predict(self.arch)
# encode object into what
self.encoder = Encoder(self.arch)
# decode what into object
self.decoder = Decoder(self.arch)
# spatial transformers
self.image_to_object = SpatialTransformer(self.arch.input_shape,
self.arch.object_shape)
self.object_to_image = SpatialTransformer(self.arch.object_shape,
self.arch.input_shape)
# baseline RNN
self.bl_rnn = LSTMCell(lstm_input_size, self.arch.baseline_hidden_size)
# predict baseline value
self.bl_predict = nn.Linear(self.arch.baseline_hidden_size, 1)
# priors
self.pres_prior = Bernoulli(probs=self.arch.z_pres_prob_prior)
self.where_prior = Normal(loc=self.arch.z_where_loc_prior,
scale=self.arch.z_where_scale_prior)
self.what_prior = Normal(loc=self.arch.z_what_loc_prior,
scale=self.arch.z_what_scale_prior)
# modules excluding baseline rnn
self.air_modules = nn.ModuleList(
[self.predict, self.lstm_cell, self.encoder, self.decoder])
self.baseline_modules = nn.ModuleList([self.bl_rnn, self.bl_predict])
def __init__(self, token_embedder, embed_dim):
super(RNNSequenceEmbedder, self).__init__()
self._embed_dim = embed_dim
word_dim = token_embedder.embed_dim
rnn_cell = LSTMCell(word_dim, embed_dim)
self.source_encoder = SimpleSourceEncoder(rnn_cell)
self.vocab = token_embedder.vocab
self.token_embedder = token_embedder
def __init__(self, token_embedder, hidden_dim, input_dim, agenda_dim):
super(SimpleDecoderCell, self).__init__()
self.rnn_cell = LSTMCell(input_dim + agenda_dim, hidden_dim)
self.linear = Linear(hidden_dim, input_dim)
self.h0 = Parameter(torch.zeros(hidden_dim))
self.c0 = Parameter(torch.zeros(hidden_dim))
self.softmax = Softmax(dim=1)
self.token_embedder = token_embedder
def init_network(self):
"""Initialize network parameters. This is an actor-critic build on top of a RNN cell. The
actor is a fully connected layer, and the critic consists of two fully connected layers"""
self.rnn = LSTMCell(self.n_actions, self._hidden_size)
for p in self.rnn.parameters():
uniform_(p, self._uniform_init[0], self._uniform_init[1])
self.actor = Linear(self._hidden_size, self.n_actions)
for p in self.actor.parameters():
uniform_(p, self._uniform_init[0], self._uniform_init[1])
self.middle_critic = Linear(self._hidden_size, self._hidden_size // 2)
for p in self.middle_critic.parameters():
uniform_(p, self._uniform_init[0], self._uniform_init[1])
self.critic = Linear(self._hidden_size // 2, 1)
for p in self.critic.parameters():
uniform_(p, self._uniform_init[0], self._uniform_init[1])
def __init__(self,
embeddings,
max_word=32,
multi_image=1,
multi_merge='att',
labels=None,
aete_s=2000,
aete_r=5,
lstm_dim=256,
lambda_a=0.85,
teacher_forcing=None,
image_model=None,
image_pretrained=None,
finetune_image=False,
image_finetune_epoch=None,
rl_opts=None,
word_idxs=None,
device='gpu',
verbose=False):
super(TieNet, self).__init__(max_word, multi_image, multi_merge,
teacher_forcing, image_finetune_epoch,
rl_opts, word_idxs, verbose)
# Label statistics
self.chexpert_labels, self.lp, self.ln, self.lq = self._load_labels(
labels)
# Various NN parameters
self.feat_dim = lstm_dim
self.lstm_dim = lstm_dim
self.lambda_a = lambda_a
self.dropout = Dropout(0.5)
# Image processes
if image_model is None:
image_model = 'resnet50'
self.image_feats, image_dim = ImageClassification.image_features(
image_model, not finetune_image, True, image_pretrained, device)
self._init_multi_image(image_dim, self.VISUAL_NUM, lstm_dim)
self.image_proj = Linear(image_dim, lstm_dim)
# Word processes
self.init_h = Linear(lstm_dim, lstm_dim)
self.init_c = Linear(lstm_dim, lstm_dim)
self.att_v = Linear(image_dim, image_dim)
self.att_h = Linear(lstm_dim, image_dim)
self.att_a = Linear(image_dim, 1)
self.gate = Linear(lstm_dim, image_dim)
input_dim = image_dim + embeddings.shape[1]
self.lstm_word = LSTMCell(input_dim, lstm_dim)
self.embeddings = Embedding.from_pretrained(
embeddings,
freeze=False,
padding_idx=PretrainedEmbeddings.INDEX_PAD)
self.embed_num = self.embeddings.num_embeddings
self.word_dense = Linear(lstm_dim, embeddings.shape[0], bias=False)
# AETE processes
self.aete1 = Linear(lstm_dim, aete_s)
self.aete2 = Linear(aete_s, aete_r)
# Joint
self.joint = Linear(lstm_dim + image_dim, self.DISEASE_NUM * 2)
def __init__(self,
input_size,
hidden_size):
super(mLSTMCell, self).__init__()
self._input_size = input_size
self._hidden_size = hidden_size
self._lstm_cell = weight_norm(LSTMCell(input_size, hidden_size), name='weight_ih')
self._lstm_cell = weight_norm(self._lstm_cell, name='weight_hh')
self._i_multiplier = weight_norm(Linear(input_size, hidden_size, bias=False))
self._h_multiplier = weight_norm(Linear(hidden_size, hidden_size, bias=False))
def __init__(self, input_size, output_size, hidden_size):
super(LSTMAuto, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.hidden_size = hidden_size
# Output of previous iteration appended to input
self.lstmCell = LSTMCell(output_size + input_size, hidden_size)
# Softmax variables
self.linear = nn.Linear(hidden_size, output_size)
self.softmax = nn.Softmax()
def __init__(self, input_shape, id, normalize, nb_hidden):
super().__init__(input_shape, id)
self._nb_hidden = nb_hidden
if normalize:
self.lstm = LSTMCellLayerNorm(input_shape[0], nb_hidden)
else:
self.lstm = LSTMCell(input_shape[0], nb_hidden)
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
def __init__(self, h_g, h_l, std, hidden_size, num_classes, patch_amount, patch_size, scale_factor):
super(AdaptiveAttention, self).__init__()
self.std = std
self.sensor = GlimpseNetwork(hidden_size, patch_amount=patch_amount, patch_size=patch_size,
scale_factor=scale_factor)
self.rnn = LSTMCell(hidden_size, hidden_size)
self.decider = DecisionNetwork(hidden_size, 2)
self.locator = LocationNetwork(hidden_size, 2, std)
self.classifier = ActionNetwork(hidden_size, num_classes)
self.baseliner = BaselineNetwork(hidden_size, 1)
def __init__(self, emb_dim, classifier_type):
"""
:param emb_dim:
:param classifier_type: тип классификатора: 1 - по состоянию верхней вершины, 2 - по состоянию всех вершин, 3 - по состоянию только вершин-переменных, 5 - вычисление по логике Заде, 6 - вычисление по вероятностной логике, 7 - вычисление по логике Лукашевича
"""
super().__init__()
self.classifier_type = classifier_type
self.emb_dim = emb_dim
self.start_embeddings1 = torch.zeros([self.emb_dim], requires_grad=False)
# self.start_embeddings = torch.nn.Parameter(torch.FloatTensor(self.emb_dim))
# torch.nn.init.normal_(self.start_embeddings, 0, 1)
self.con_embeddings1 = torch.tensor([1., 0., 0., 0., ], requires_grad=False)
# self.con_embeddings = torch.nn.Parameter(torch.FloatTensor(self.emb_dim))
# torch.nn.init.normal_(self.con_embeddings, 0, 1)
self.dis_embeddings1 = torch.tensor([0., 1., 0., 0., ], requires_grad=False)
# self.con_embeddings = torch.nn.Parameter(torch.FloatTensor(self.emb_dim))
# torch.nn.init.normal_(self.con_embeddings, 0, 1)
self.neg_embeddings1 = torch.tensor([0., 0., 1., 0., ], requires_grad=False)
# self.neg_embeddings = torch.nn.Parameter(torch.FloatTensor(self.emb_dim))
# torch.nn.init.normal_(self.neg_embeddings, 0, 1)
self.var_embeddings1 = torch.tensor([0., 0., 0., 1., ], requires_grad=False)
# self.var_embeddings = torch.nn.Parameter(torch.FloatTensor(self.emb_dim))
# torch.nn.init.normal_(self.var_embeddings, 0, 1)
# self.msg1_1_f = Linear(2 * self.emb_dim, 2 * self.emb_dim, True)
self.msg1_1_f = Linear(self.emb_dim + 4, 2 * self.emb_dim, True)
self.msg1_2_f = Linear(2 * self.emb_dim, self.emb_dim, True)
# self.msg2_1_f = Linear(2 * self.emb_dim, 2 * self.emb_dim, True)
self.msg2_1_f = Linear(self.emb_dim + 4, 2 * self.emb_dim, True)
self.msg2_2_f = Linear(2 * self.emb_dim, self.emb_dim, True)
self.update_f = LSTMCell(self.emb_dim, self.emb_dim, self.emb_dim)
self.clf_1 = Linear(self.emb_dim, self.emb_dim, True)
self.clf_2 = Linear(self.emb_dim, 1, False)
if self.classifier_type in [1, 2, 3, 4]:
pass
# self.cl = Linear(self.emb_dim, 1, False)
elif self.classifier_type == 5:
# self.tvalue = Linear(self.emb_dim, 1, False)
self.cl = GodelEvaluation()
elif self.classifier_type == 6:
# self.tvalue = Linear(self.emb_dim, 1, False)
self.cl = ProbabilisticEvaluation()
elif self.classifier_type == 7:
# self.tvalue = Linear(self.emb_dim, 1, False)
self.cl = LukasieviczEvaluation()
def __init__(self,
input_size: int,
hidden_size: int = 200,
bias: bool = True,
dropout: float = 0.):
super().__init__(hidden_size)
self.output_dim = hidden_size
self.states: List[Tensor] = list()
self.items = list()
self.cell = LSTMCell(input_size, hidden_size, bias)
self.dropout = Dropout(dropout) # TODO:每一个instance的mask要一致
def __init__(self, nb_input_channel, nb_out_channel, normalize):
super().__init__()
self._nb_output_channel = nb_out_channel
if normalize:
self.lstm = LSTMCellLayerNorm(nb_input_channel,
self._nb_output_channel)
else:
self.lstm = LSTMCell(nb_input_channel, self._nb_output_channel)
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
def __init__(self,params:configargparse.Namespace,att: torch.nn.Module=None):
"""
Neural Network Module for the Sequence to Sequence LAS Model
:params configargparse.Namespace params: The training options
:params torch.nn.Module att: The attention module
"""
super(Speller,self).__init__()
## Embedding Layer
self.embed = Embedding(params.odim,params.demb_dim)
## Decoder with LSTM Cells
self.decoder = ModuleList()
self.dropout_dec = ModuleList()
self.dtype = params.dtype
self.dunits = params.dhiddens
self.dlayers = params.dlayers
self.decoder += [
LSTMCell(params.eprojs + params.demb_dim, params.dhiddens)
if self.dtype == "lstm"
else GRUCell(params.eprojs + params.demb_dim, params.dhiddens)
]
self.dropout_dec += [Dropout(p=params.ddropout)]
self.dropout_emb = Dropout(p=params.ddropout)
## Other decoder layers if > 1 decoder layer
for i in range(1,params.dlayers):
self.decoder += [
LSTMCell(params.dhiddens, params.dhiddens)
if self.dtype == "lstm"
else GRUCell(params.dhiddens, params.dhiddens)
]
self.dropout_dec += [LockedDropout(p=params.ddropout)] # Dropout
## Project to Softmax Space- Output
self.projections = Linear(params.dhiddens, params.odim)
## Attention Module
self.att = att
## Scheduled Sampling
self.sampling_probability = params.ssprob
## Initialize EOS, SOS
self.eos = len(params.char_list) -1
self.sos = self.eos
self.ignore_id = params.text_pad
def __init__(self,
input_dim: int,
hidden_dim: int,
num_layers: int = 1,
layer_dropout: float = 0.0,
recurrent_dropout: float = 0.0):
super().__init__(input_dim, hidden_dim)
self.hidden = None
self.context = None
assert num_layers >= 1
self.layers = num_layers
self._lstm_cell0 = LSTMCell(input_dim, hidden_dim)
self._lstm_cellL = [
LSTMCell(hidden_dim, hidden_dim) for _ in range(num_layers - 1)
]
self.layer_dropout_rate = layer_dropout
self.recurrent_dropout_rate = recurrent_dropout
if self.layer_dropout_rate > 0.0 and num_layers < 2:
raise ConfigurationError(
"Layer dropout must be 0.0 if we have only a single layer")
def __init__(self,
encoder,
hidden_size,
num_programs,
num_non_primary_programs,
embedding_dim,
encoding_dim,
indices_non_primary_programs,
learning_rate=1e-3,
temperature=0.1):
super(Policy, self).__init__()
self._uniform_init = (-0.1, 0.1)
self._hidden_size = hidden_size
self.num_programs = num_programs
self.num_non_primary_programs = num_non_primary_programs
self.embedding_dim = embedding_dim
self.encoding_dim = encoding_dim
# Initialize networks
self.Mprog = Embedding(num_non_primary_programs, embedding_dim)
self.encoder = encoder
self.lstm = LSTMCell(self.encoding_dim + self.embedding_dim,
self._hidden_size)
self.critic = CriticNet(self._hidden_size)
self.actor = ContinuousActorNet(self._hidden_size, self.num_programs)
self.temperature = temperature
self.init_networks()
self.init_optimizer(lr=learning_rate)
# Compute relative indices of non primary programs (to deal with task indices)
self.relative_indices = dict(
(prog_idx, relat_idx)
for relat_idx, prog_idx in enumerate(indices_non_primary_programs))
def init_network(self):
"""Initialize network parameters. This is an actor-critic build on top of a RNN cell. The
actor is a fully connected layer, and the critic consists of two fully connected layers"""
self.rnn = LSTMCell(self.action_space, self.hidden_size)
for p in self.rnn.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.actor = Linear(self.hidden_size, self.action_space)
for p in self.actor.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.middle_critic = Linear(self.hidden_size, self.hidden_size // 2)
for p in self.middle_critic.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.critic = Linear(self.hidden_size // 2, 1)
for p in self.critic.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.encoder = resnet34(**{"num_classes": self.embedding})
self.padding = ZeroPad2d((30, 20, 0, 0))
def __init__(
self,
decoding_dim: int,
target_embedding_dim: int,
attention: Optional[Attention] = None,
bidirectional_input: bool = False,
num_decoder_layers: int = 1,
accumulate_hidden_states: bool = False,
dropout: float = 0.2,
) -> None:
super().__init__(
decoding_dim=decoding_dim,
target_embedding_dim=target_embedding_dim,
decodes_parallel=False,
)
# In this particular type of decoder output of previous step passes directly to the input of current step
# We also assume that decoder output dimensionality is equal to the encoder output dimensionality
decoder_input_dim = self.target_embedding_dim
# Attention mechanism applied to the encoder output for each step.
self._attention = attention
if self._attention:
# If using attention, a weighted average over encoder outputs will be concatenated
# to the previous target embedding to form the input to the decoder at each
# time step. encoder output dim will be same as decoding_dim
decoder_input_dim += decoding_dim
# Ensure that attention is only set during seq2seq setting.
# if not self._seq2seq_mode and self._attention is not None:
# raise ConfigurationError("Attention is only specified in Seq2Seq setting.")
self._num_decoder_layers = num_decoder_layers
if self._num_decoder_layers > 1:
self._decoder_cell = LSTM(
input_size=decoder_input_dim,
hidden_size=self.decoding_dim,
num_layers=self._num_decoder_layers,
dropout=dropout,
)
else:
# We'll use an LSTM cell as the recurrent cell that produces a hidden state
# for the decoder at each time step.
# TODO (pradeep): Do not hardcode decoder cell type.
self._decoder_cell = LSTMCell(decoder_input_dim, self.decoding_dim)
self._bidirectional_input = bidirectional_input
self._accumulate_hidden_states = accumulate_hidden_states
def __init__(self,
embeddings,
max_word=32,
multi_image=1,
multi_merge='att',
context_dim=512,
lstm_dim=1000,
lambda_a=1.0,
teacher_forcing=None,
image_model=None,
image_pretrained=None,
finetune_image=False,
image_finetune_epoch=None,
rl_opts=None,
word_idxs=None,
device='gpu',
verbose=False):
super(ShowAttendAndTell,
self).__init__(max_word, multi_image, multi_merge,
teacher_forcing, image_finetune_epoch, rl_opts,
word_idxs, verbose)
self.feat_dim = context_dim
self.lstm_dim = lstm_dim
self.lambda_a = lambda_a
self.dropout = Dropout(0.5)
# Image processes
if image_model is None:
image_model = 'vgg'
self.image_feats, image_dim = ImageClassification.image_features(
image_model, not finetune_image, True, image_pretrained, device)
self._init_multi_image(image_dim, self.VISUAL_NUM, lstm_dim)
self.image_proj = Linear(image_dim, context_dim)
# Word processes
self.init_h = Linear(context_dim, lstm_dim)
self.init_c = Linear(context_dim, lstm_dim)
self.att_v = Linear(image_dim, image_dim)
self.att_h = Linear(lstm_dim, image_dim)
self.att_a = Linear(image_dim, 1)
self.gate = Linear(lstm_dim, image_dim)
input_dim = image_dim + embeddings.shape[1]
self.lstm_word = LSTMCell(input_dim, lstm_dim)
self.embeddings = Embedding.from_pretrained(
embeddings,
freeze=False,
padding_idx=PretrainedEmbeddings.INDEX_PAD)
self.embed_num = self.embeddings.num_embeddings
# Deep output
self.lh = Linear(lstm_dim, embeddings.shape[1])
self.lz = Linear(image_dim, embeddings.shape[1])
self.lo = Linear(embeddings.shape[1], embeddings.shape[0], bias=False)
def __init__(self, target_token_embedder, input_dim, agenda_dim,
decoder_dim, encoder_dim, attn_dim, num_layers, num_inputs,
dropout_prob, disable_attention):
super(AttentionDecoderCell, self).__init__()
target_dim = target_token_embedder.embed_dim
self.num_layers = num_layers
self.num_inputs = num_inputs
self.disable_attention = disable_attention
if disable_attention:
augment_dim = agenda_dim
else:
# see definition of `x_augment` in `forward` method
# we augment the input to each RNN layer with num_inputs attention contexts + the agenda
augment_dim = encoder_dim * num_inputs + agenda_dim
self.rnn_cells = []
for layer in range(num_layers):
in_dim = input_dim if layer == 0 else decoder_dim # first layer takes word vectors
out_dim = decoder_dim
rnn_cell = LSTMCell(in_dim + augment_dim, out_dim)
self.add_module('decoder_layer_{}'.format(layer), rnn_cell)
self.rnn_cells.append(rnn_cell)
if disable_attention:
z_dim = decoder_dim
else:
# see definition of `z` in `forward` method
# to predict words, we condition on the hidden state h + num_inputs attention context
z_dim = decoder_dim + encoder_dim * num_inputs
# TODO(kelvin): these big params may need regularization
self.vocab_projection_pos = Linear(z_dim, target_dim)
self.vocab_projection_neg = Linear(z_dim, target_dim)
self.relu = torch.nn.ReLU()
self.h0 = Parameter(torch.zeros(decoder_dim))
self.c0 = Parameter(torch.zeros(decoder_dim))
self.vocab_softmax = Softmax()
self.input_attentions = []
for i in range(num_inputs):
attn = Attention(encoder_dim, decoder_dim, attn_dim)
self.add_module('input_attention_{}'.format(i), attn)
self.input_attentions.append(attn)
self.target_token_embedder = target_token_embedder
self.dropout = Dropout(dropout_prob)
def __init__(self, hparams):
super().__init__()
self.input_size = (1 + hparams.position_number) * hparams.embedding_dim
self.hidden_size = hparams.argument_size * hparams.relation_size * hparams.position_size
self.zeroth_tuple = Parameter(empty(1, self.input_size))
self.zeroth_cell_state = Parameter(empty(1, self.hidden_size).zero_())
self.cell = LSTMCell(self.input_size, self.hidden_size)
self.attention = Attention(hparams)
self.unbinding_module = UnbindingModule(hparams)
self.init_weights()
def __init__(self, output_size, hidden_size, seq_len, num_layers=1, bias=True, dropout=0, bidirectional=False):
super(DecoderLSTM, self).__init__()
self.output_size = output_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.bias = bias
self.dropout = dropout
self.dropout_state = {}
self.bidirectional = bidirectional
self.seq_len = seq_len
num_directions = 2 if bidirectional else 1
self.lstm = LSTMCell(output_size, hidden_size)
self.linear = nn.Linear(hidden_size, output_size)
self.softmax = nn.Softmax()
def __init__(self, input_size, output_size, hidden_size, num_layers, lookup):
super(LSTMAutoParams, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.lookup = lookup
# Output of previous iteration appended to input
self.layers = []
for i in range(num_layers):
self.layers.append(LSTMCell(input_size, hidden_size))
input_size = hidden_size
# Softmax variables
self.linear = nn.Linear(hidden_size, output_size)
self.softmax = nn.Softmax()
def __init__(self, nb_input_channel, nb_out_channel, normalize):
super().__init__()
self._nb_output_channel = 256
self.linear = Linear(2592, self._nb_output_channel)
if normalize:
self.lstm = LSTMCellLayerNorm(
self._nb_output_channel,
self._nb_output_channel) # hack for experiment
self.bn_linear = BatchNorm1d(self._nb_output_channel)
else:
self.bn_linear = Identity()
self.lstm = LSTMCell(self._nb_output_channel,
self._nb_output_channel)
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
